Deep learning-based classification of infectious keratitis on slit-lamp images

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Infectious keratitis (IK) is a critical eye condition requiring prompt diagnosis.
• Traditional culture-based methods for diagnosing IK have limitations including delayed results and low sensitivity.
• There is a need for advanced diagnostic tools to aid in the early and accurate identification of IK.

Purpose of the Study:

• To develop and validate a deep learning-based auxiliary diagnostic model for the early detection of infectious keratitis.
• To assess the performance of the deep learning model in classifying different types of microbial keratitis.

Main Methods:

• A retrospective analysis of 4830 slit-lamp images from patients with confirmed infectious keratitis (bacterial, fungal, Acanthamoeba, herpes simplex).
• Implementation and comparison of five image classification networks, with a blended model (KeratitisNet) combining ResNext101_32x16d and DenseNet169.
• Validation using 10-fold cross-validation, ROC curves, confusion matrices, and comparison against diagnoses by three experienced cornea specialists.

Main Results:

• KeratitisNet achieved an overall accuracy of 77.08% in diagnosing infectious keratitis.
• Specific accuracies for bacterial keratitis (BK), fungal keratitis (FK), Acanthamoeba keratitis (AK), and herpes simplex keratitis (HSK) were 70.27%, 77.71%, 83.81%, and 79.31%, respectively.
• The model's diagnostic accuracy for each keratitis type was significantly higher than that of human ophthalmologists (p < 0.001).

Conclusions:

• The developed deep learning model, KeratitisNet, demonstrates strong performance in diagnosing and classifying infectious keratitis.
• Deep learning offers a valuable auxiliary tool for clinicians, assisting in the suspicion of IK based on corneal manifestations.
• AI-powered diagnostic systems can enhance the early detection and management of infectious keratitis.